Mistral AI has released Voxtral 3, an integrated suite of language models engineered specifically for building real-time voice conversational agents. The modular stack features Voxtral Small Transcribe 3 for speech recognition, Voxtral Mini TTS 3 for speech synthesis, and a dedicated LiveKit-based software development kit. Together, these components unify listening, reasoning, external tool usage, and speech delivery into a single continuous streaming pipeline.
The system offers broad regional coverage, supporting transcription across more than 50 languages alongside text-to-speech generation in over 35 languages and 50 regional dialects with emotional adaptation. Voice agents created with the platform can interact with external business tools during live dialogue, permitting automated workflows such as appointment scheduling or account invoice lookup.
Mistral is offering the stack via self-hosted deployments on customer infrastructure as well as hosted APIs within the European Union to meet stringent data sovereignty compliance. Commercial pricing for the transcription layer starts at $0.0035 per minute, establishing a baseline for enterprise voice automation.
Why it matters
Developers can construct low-latency, multilingual voice agents capable of executing live business tools within custom or EU-compliant infrastructure.
Enterprise buyers gain an alternative to US cloud providers for sovereign voice AI deployments across dynamic customer service applications.
Source: the-decoder.com



